Skip to content
All library documents

Why Six Months of Intraday Data Is Likely Too Short for Volatility Forecasting

Article Quant Q&A · Author: opt
Read the original

Summary

The document considers whether roughly six months of five-minute observations are enough to compare HAR-RV, realized GARCH, and HEAVY models for out-of-sample realized-volatility forecasting. Its answer focuses on the number of trading days represented, estimating about 132 days when assuming 22 sessions per month. It judges that sample likely too small for reliable forecasting and cautions against expecting a meaningful model comparison from so little history.

The response supports that caution by pointing to a volatility-modeling survey whose authors used roughly ten years of data. This is a comparison, not a formal minimum-sample result or a test of the three models on the questioner's data. It does not report model estimates, forecast errors, or a winner. The practical lesson is to treat results from a short sample as highly uncertain and seek substantially longer history where possible; the answer does not discuss intraday sampling quality, market regimes, or how model performance might vary by asset.

Key ideas

  • The answer estimates that six months of observations amount to about 132 trading days under its stated calendar assumption.
  • It considers that history too short to support confident volatility forecasting.
  • A cited survey used a much longer data span, offering context rather than a formal minimum requirement.
  • The document provides no empirical comparison or ranking of HAR-RV, realized GARCH, and HEAVY.

Tags

VolatilityStatisticsBacktesting
Full text
# HAR-RV, realized GARCH and HEAVY model for realized volatility


# HAR-RV, realized GARCH and HEAVY model for realized volatility












I don't have much experience with volatility modeling using intraday data but I'm in the process of collecting 5mins data. Currently I have ~6 months of data. Is it enough to use these models with such short history? Which one should perform better out of sample having this small amount of data?

## Answer by user11953 (score 1, accepted)

https://quant.stackexchange.com/a/14595

Concidering 22 days of trading per month you have approximatly 132 days of trading. I highly doubt that this will be sufficient for any forecasting. The sample might be too small. Have a look here: http://research.stlouisfed.org/wp/2012/2012-008.pdf Erdemlioglu, Laurent and Neely used the data of ~10 years to conduct their survey.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.